Per-separation clustered-dot color halftone watermarks: separation estimation based …oztanb/Papers/jei10paper.pdf · 2010-12-17 · Journal of Electronic Imaging 19(4), 043007 (Oct–Dec

Journal of Electronic Imaging 19(4), 043007 (Oct–Dec 2010)

Per-separation clustered-dot color halftone watermarks:separation estimation based on spatial frequency content

Basak Oztan∗University of Rochester

Department of Electrical and Computer EngineeringRochester, New York 14627-0126

Gaurav SharmaUniversity of Rochester

Department of Electrical and Computer Engineeringand

Department of Biostatistics and Computational BiologyRochester, New York 14627-0126

E-mail: [email protected]

Abstract. A framework for clustered-dot color halftone watermarkingis considered, wherein watermark patterns are embedded in individ-ual colorant halftones prior to printing and embedded watermarksare detected from scans of the printed images after obtaining esti-mates of the individual halftone separations. The principal challengein this methodology arises in the watermark detection phase. Typicalthree-channel RGB scanner systems do not directly provide good es-timates of the four CMYK colorant halftones that are commonly usedin color printing systems. To address this challenge, we propose anestimation method that, when used with suitably selected halftoneperiodicities, jointly exploits the differences in the spatial periodicitiesand the color (spectra) of the halftone separations to obtain goodestimates of the individual halftones from conventional RGB scans.We demonstrate the efficacy of this methodology experimentally us-ing continuous phase modulation for the embedding of independentvisual watermark patterns in the individual halftone separations. Wa-termarks detected from the estimates of halftone separations ob-tained using the proposed estimation method have a much highercontrast than those detected directly. We also evaluate the accuracyof the estimated halftones through simulations and demonstrate thatthe proposed estimation method offers high accuracy. © 2010 SPIEand IS&T. [DOI: 10.1117/1.3497615]

1 IntroductionDigital watermarks have recently emerged as an importantenabling technology for security and forensics applicationsfor multimedia3–6 and for hardcopy (Ref. 7, Chap. 5, andRef. 8). In the hardcopy domain these techniques provide

∗Currently with the Department of Computer Science, Rensselaer Poly-technic Institute, Troy, New York 12180.

This paper is available online as an open-access article, with color versionsof several of the figures. In particular, Figs. 1, 2, 4 and 29 may be difficultto interpret without color; print readers should refer to the online version athttp://SPIEDigitalLibrary.org for these figures.

Paper 10002PRR received Jan. 5, 2010; revised manuscript received Aug.11, 2010; accepted for publication Aug. 24, 2010; published online Dec. 16,2010

1017-9909/2010/19(4)/043007/22/$25.00 C© 2010 SPIE and IS&T.

functionality that mimics or extends the capabilities of con-ventional paper watermarks that have been extensively uti-lized since their introduction in the late thirteenth century.9

Since most hardcopy reproduction relies on halftone print-ing, methods that embed the watermark in the halftone struc-ture comprise one of the primary categories of hardcopydigital watermarks. These methods enable printed images tocarry watermark data in the form of changes in the halftonestructures, which are normally imperceptible but can be dis-tinguished by appropriate detection methods. The proposedmethods for detection of halftone watermarks can involveeither scanning and digital processing or manual detection,wherein the ordinarily imperceptible watermark pattern isrendered visible by overlaying the printed image with a de-coder mask such as a preprinted transparency. Because themanual process can also be simulated in the digital process-ing, manually detectable watermarks can typically also bedetected via digital processing, whereas the converse doesnot necessarily hold. A number of techniques have been pro-posed for halftone-based watermarking in black and white,i.e., monochrome, printing systems.10–20 Most of these meth-ods, however, do not directly generalize to color. Thus, thereis an unmet need for color halftone watermarking techniques.This need is exacerbated by the fact that image content is of-ten printed in color, particularly as color printing systemsbecome more affordable and accessible.

Since the colorant halftone separations are printed sequen-tially and overlaid on the paper substrate, per-channel em-bedding and detection of watermark patterns is an attractiveoption that would extend monochrome watermarking meth-ods to color. To detect the watermark patterns, one would liketo acquire the constituent halftone separations used in thecolor printing system from the (overlaid) print. This wouldwork, for example, if one could deploy a scanner with Ncolor channels, where each color channel captures only oneof the N colorants used in the printing system. In actual prac-tice, desktop scanners commonly use RGB color filters tocapture color. For three-color CMY printing, as illustrated in

Journal of Electronic Imaging Oct–Dec 2010/Vol. 19(4)043007-1

Downloaded from SPIE Digital Library on 17 Dec 2010 to 129.161.33.116. Terms of Use: http://spiedl.org/terms

mailto: [email protected]

mailto: http://SPIEDigitalLibrary.org

Oztan and Sharma: Per-separation clustered-dot color halftone watermarks: separation estimation . . .

1.0

600500400

ReflectancePrint

Y increasing

M increasing

C increasing

700

Wavelength (λ)

600500400 700

Wavelength (λ)

ScannerSensitivity

RB G

Fig. 1 Complementary relation between printer CMY colorants andscanner RGB channels. (Color online only.)

Fig. 1, there is a complementary relationship between theCMY colorants and RGB scanner channels: the cyan (C),magenta (M), and yellow (Y) colorants absorb light, respec-tively, over the spectral regions in which the red (R), green(G), and blue (B) scanner channels are sensitive. Thus, inan ideal setting, C, M, and Y colorant halftones may be esti-mated from the scanner R, G, and B channels, respectively. Inpractice, however, this is usually not feasible for a couple ofreasons. First, typical printing systems utilize CMYK (four)colorants. The black (K) colorant absorbs uniformly acrossthe spectrum, and thus it consistently appears in the scannerRGB channels. Second, so-called “unwanted absorptions” ofthe CMY colorants also cause cross-coupling, i.e., C, M, andY halftone separations not only appear in the scan R, G, andB channels that complement their spectral absorption bands,

respectively, but also in the two other channels as well. Theimage of Fig. 2 illustrates the couplings between the CMYKhalftone separations in the scanner RGB channels. Due tothese undesired couplings, the four C, M, Y, and K halftoneseparations cannot be directly obtained from individual R, G,and B scanner channel responses.

In this paper, we present a methodology for estimatingindividual C, M, Y, and K halftones from RGB scans ofprinted images by jointly exploiting the differences betweensuitably selected spatial halftone periodicities and thedifferences in the spectral characteristics of the colorants.We use the proposed joint estimation scheme to extend, tocolor halftone printing, the halftone watermarking methodusing continuous phase modulation21 (CPM), which embedsand detects a visual watermark pattern. Our experimentsdemonstrate that the estimated halftones obtained via theproposed method improve the detection of the embeddedCPM watermarks, offering much higher contrast for thedetected watermark patterns than direct detection. Becausethe estimation methodology works only with electronicprocessing, the proposed method is not applicable to directdetection of the watermark in the printed image by physicaloverlay using a suitable decoder mask.

The framework we present may also be independentlyviewed as a general method for estimating CMYK halftoneseparations from RGB scans. We evaluate the accuracy ofthe proposed method in this context via simulations andcompare against alternatives that use either the differencein spatial periodicities or in spectral characteristics alone(for the estimation of the halftones). Our results demonstratethat the combined methodology offers a significant advant-age.

The presented work builds on and extends our priorwork in this area1 that addressed the more limited settingof three-color CMY printing and improves on our prelimi-nary report on this effort2 by utilizing a suitable model foranalysis and processing. Note that our framework relies ondata embedding in the individual spatial patterns of the con-stituent CMYK halftones. This is complementary to alternatemethods22–26 that have previously been proposed for datahiding in printed color images using spectral characteristicsalone. Also, dot-on-dot27 color halftone watermarking withthe same watermark embedded in each colorant channelcan be viewed as an alternative, though rather restrictive,framework for extending monochrome halftone watermark-ing methods to color.28

(a) (b) (c) (d)

Fig. 2 RGB scan of a printed image that was divided into four stripes, where the stripes from left toright contain only C, M, Y , and K colorants, respectively. The K colorant can be consistently observedin all scan channels and cross-coupling between C, M, and Y colorant halftone separations can alsobe clearly seen in the scan R, G, and B channels. (a) RGB scan of CMYK halftone; (b) scan R channel;(c) scan G channel, and (d) scan B channel. (Color online only.)




wK

Watermark

IC(x, y)

wC &

EmbeddingWatermark

Halftoning

Halftoning

WatermarkEmbedding

&

&

EmbeddingWatermark

Halftoning

IhY (x, y)wY

IY (x, y)

Halftoning

WatermarkEmbedding

&

&Print

Scan

WatermarkExtraction

ExtractionWatermark

WatermarkExtraction

Spatial Frequency

and

Color

Separation

IhY (x, y) wY

IhK(x, y)

IhM(x, y)

IhC(x, y)IhC(x, y)

IM(x, y)

wM

IhM(x, y)

wK

IK(x, y)IhK(x, y)

IC,M,Y,K(x, y) IsR,G,B (x,y) IsG(x, y)IhC, M, Y, K(x, y)

IsR(x, y)

IsB(x, y)

wM

wC

Extraction

Fig. 3 Overview of per-channel basis clustered-dot color halftone watermarking by exploiting spatialfrequency and color separation methodology.

The rest of the paper is organized as follows. Section 2describes our framework for clustered-dot color halftonewatermarking. Section 3 describes our method for estimat-ing CMYK halftone separations from RGB scans exploitingthe differences in spatial frequency and colorant spectra. Thewatermarking technique used to test our framework is de-scribed in Sec. 4, and the experimental results are presentedin Sec. 5. Finally, in Sec. 6, we present conclusions.

2 Clustered-Dot Color Halftone WatermarkingFor typical CMYK printing, the overall watermark embed-ding and extraction scheme that we consider is illustratedin Fig. 3. The input cover image IC,M,Y,K (x, y) is a con-tone CMYK image that is typically obtained by transform-ing a device-independent colorimetric representation of theimage to the device-dependent CMYK values via a set ofcolor conversions.29 The watermark wi for the i’th colorantseparation Ii (x, y), where i is one of C, M, Y, or K, is em-bedded in the halftone separation I h

i (x, y) in the halftoningstage. For the method we demonstrate, the watermark wiwill in fact be a spatial pattern wi (x, y) corresponding, forinstance, to rasterized text or a logo. The color halftone imageI hC,M,Y,K (x, y) is obtained by printing the constituent halftone

separations in overlay. For the detection of the watermarkpatterns embedded in these separations, a scan I s

R,G,B(x, y)of the printed image is obtained using a conventional RGBscanner.

To obtain estimates of the four constituent halftone sepa-rations from these three channels, additional information isrequired. Here, we rely on the spatial and spectral charac-teristics of rotated clustered-dot color halftones27 to providea solution to this problem. If the halftone frequencies aresuitably chosen, the differences in spatial periodicities andcolorant spectra can be jointly exploited to obtain estimatesof the C, M, Y, and K halftone separations from the R, G, and Bchannels I s

k (x, y), k ∈ {R, G, B}, of the scanned image—asis described in Sec. 3. Once the estimate I h

i (x, y) is obtainedfor the i th colorant halftone, the monochrome watermarkdetection method can be applied to the separation in or-der to recover (an estimate of) the corresponding watermarkwi .

3 Halftone Separation Estimation ExploitingDifferences in Spatial Periodicity andColorant Spectra

Our goal is to obtain estimates of the CMYK halftone sepa-rations I h

i (x, y), i ∈ {C, M, Y, K }, from the RGB channelsof the scanned image I s

k (x, y), k ∈ {R, G, B}. The halftonesI hi (x, y) and their estimates I h

i (x, y), i ∈ {C, M, Y, K }, arebinary images, where we adopt the convention that the values1 and 0 correspond, respectively, to whether ink/toner i is, oris not, deposited (or estimated to be deposited) at the pixelposition (x, y). On the other hand, the kth color channel ofthe scanner I s

k (x, y) is a contone image, where the imagevalue at pixel position (x, y) represents the fraction of thelight reflected from the pixel (x, y) within the transmittanceband of the kth scanner channel.

As we pointed out earlier, for typical desktop RGB scan-ners, coupling between the different colorant halftones in thescanned RGB image is inevitable. Figures 4(a) to 4(d) showenlarged views of a region from the digital M halftone sep-aration, digital CMYK halftone separation overlay, scannedRGB image, and scan G channel of an image used in our ex-periments to test our framework, respectively. It is apparentnot only is the halftone structure of the desired M separationvisible, but due to the fact that the C, Y, and K separationsalso have absorption in the scanner G channel, the undesiredhalftone structure for these separations can be seen in the Gchannel with varying intensities. This interference can be alsoobserved in the frequency domain. Figure 5 shows enlargedview of the log-magnitude Fourier spectrum of the G chan-nel image for the experimental data around low-frequencyregions. The Fourier spectrum exhibits peaks not only aboutthe fundamental frequency vectors (and higher order har-monics) of the M separation, but also at the fundamentalfrequency vectors of C, Y, and K separations, and moire fre-quencies. We exploit the fact that, when suitably designed,these peaks appear at distant locations in the frequency do-main and unwanted frequency components can be eliminatedby spatial filtering.

A schematic overview of our color halftone separationestimation algorithm is shown in Fig. 6. First, we obtain,via spatial filtering, an estimate of (black) K halftone sepa-ration using the scan RGB channel images jointly. Detailed




(a) (b) (c) (d) (e)

Fig. 4 (a) Enlarged view of a region from the digital M halftone separation of “Hats” halftone image usedin our experiments Ih

M (x, y), (b) the same region from the digital CMYK halftone overlay IhC,M,Y,K(x, y),

(c) approximately the same region from the RGB scanned color halftone print IsR,G,B(x, y), (d) the region

from the G channel image of the scanned print showing the couplings between the halftone structuresof M and C, Y , and K separations Is

G(x, y), and (e) the region from the M halftone estimate obtainedusing our methodology Ih

M (x, y). (Color online only.)

description of this process is presented in Sec. 3.2. The es-timate of the K halftone separation is then utilized togetherwith each of the scanner R, G, and B channel images to ob-tain estimates of the complementary C, M, and Y halftoneseparations, respectively, where once again, the spatial andspectral information is used jointly. The estimation processof CMY halftone separations is described in detail in Sec. 3.3and a sample estimate obtained from the scanned image ofFig. 4(d) for the magenta separation, for approximately thesame regions as Fig. 4(a), is shown in Fig. 4(e). For bet-ter understanding of the method and its limitations, we firstpresent a joint spatial-spectral analysis for clustered-dot colorhalftones and also carry through a 1-D example for whichthe results are easy to interpret.

u (lpi)

v(lpi

)

(fC , φC

)↘

↙(fM , φM

)↙

(fY , φY

)

↑(fK , φK

)

−100 −50 0 50 100

100

50

0

−50

100

Fig. 5 Enlarged view of the log-magnitude Fourier spectrum of the Gchannel of the scanned “Hats” halftone print used in our experiments.For illustration purposes, the frequencies of the constituent halftoneseparations are indicated by arrow and text labels.

3.1 Fourier Analysis of Clustered-Dot ColorHalftones

We analyze a general setting where the periodicities of the in-dividual colorant halftones are defined by 2-D lattice tilingsof the plane.30 The lattice �i for the i th colorant separa-tion is defined as �i = {Vi n | n ∈ Z

2}, where Vi is a 2×2real-valued matrix, whose columns are the two linearly in-dependent basis vectors that represent the periodicity of thehalftone separation, n = [nx ny]T , and Z denotes the set ofintegers. We first investigated the case where the contoneimage for the i th colorant, i ∈ {C, M, Y, K }, is a constantgray-level image Ii (x, y) = Ii , where x and y represent thespatial coordinates along the horizontal and vertical direc-tions, respectively. In this case, the corresponding halftoneimage corresponds to shifted replicates of a single spot func-tion si (x, y; Ii ) placed at each point in �i (Refs. 31 to 33). Asthis halftone is periodic on the lattice, it can be representedin terms of a Fourier series31–33 as

I hi (x, y) =

∞∑nx =−∞

∞∑ny=−∞

cIinx ,ny

exp [ j2π (Wi n)T x], (1)

where

cIinx ,ny

= 1

|Ui |∫ ∫

Ui

si (x, y; Ii ) exp [− j2π (Wi n)T x]dxdy,

(2)

are the Fourier series coefficients for the halftone spot func-tion si (x, y; Ii ), the matrix Wi = (V−1

i )T includes the twolinearly independent fundamental screen frequency vectors

IhK(x, y)

EstimationM Halftone

Y HalftoneEstimation

C HalftoneEstimation

IhY (x, y)

IhM (x, y)

IhC(x, y)

IsG(x, y)

IsB(x, y)

IsR(x, y)

Estimation

K Halftone

IsR(x, y)

IsB(x, y)

IsG(x, y)

Fig. 6 Overview of CMYK halftone separation estimation using theRGB channels of the scanned image.




Vi = [vi1 vi

2]

vi1

...

· · ·8

8

16

16· · ·

...

21

21 x

y

vi2

1

2

wi2

wi1

Wi = [wi1 wi

2]

u

v

1 2

(a) (b)

Fig. 7 (a) Example of a halftone periodicity lattice �i and (b) the corresponding reciprocal spatial frequency lattice �∗i .

for the i th halftone separation as its columns, Ui is the unitcell for the lattice �i and |Ui | is its area, n = [nx ny]T

and x = [x y]T . Recall, that in our notational conventionhalftone image I h

i (x, y) values of 1 and 0 correspond, re-spectively, to regions with and without the i th colorant. Thelattice �∗

i = {Wi n | n ∈ Z2} constitutes the reciprocal spa-

tial frequency lattice for the lattice �i . The Fourier spectrumof I h

i (x, y) is comprised of impulses at frequencies u ∈ �∗i ,

that correspond to the Fourier spectrum of the halftone spotfunction si (x, y; Ii ) sampled at u = Wi n, where u = [u, v]T ,and u and v represent the frequency coordinates along thehorizontal and vertical directions, respectively. To aid under-standing, we include in Fig. 7 an illustration of a halftoneperiodicity lattice �i and its reciprocal spatial frequencylattice �∗

i , where the basis vectors corresponding to thecolumns of Vi and Wi are also indicated. Additional illustra-tions and mathematical details of the representation can befound in Refs. 8, 33, and 34.

A similar representation can be obtained for the halftoneimage for a contone image with varying image content bydecomposing the contone image in the form of a siftingintegral and combining it with the Fourier series representa-tion for the halftone for a constant gray level image as35, 36

I hi (x, y) =

∞∑nx =−∞

∞∑ny=−∞

cIi (x,y)nx ,ny

exp [ j2π (Wi n)T x], (3)

where

cIi (x,y)nx ,ny

= 1

| Ui |∫ ∫

Ui

si [x, y; Ii (x, y)]

× exp [− j2π (Wi n)T x]dx dy (4)

are the Fourier series coefficients for the halftone spot func-tion si [x, y; Ii (x, y)] for the image gray-level Ii (x, y). Thefunctions cIi (x,y)

nx ,ny have spatial frequency bandwidths that arerather small as compared to the frequencies Wi n, n ∈ Z

2. TheFourier spectrum of the halftone image obtained by a Fouriertransform of the preceding identity therefore includes thespectrum of the original image (centered at dc) and the spec-tra of nonlinearly transformed versions cIi (x,y)

nx ,ny of the original

image generated by the halftoning, each shifted to a corre-sponding frequency u = Wi n, n ∈ Z

2 (Refs. 35 and 36).Having described a model for the individual halftone sep-

arations, we now consider an idealized model for the printingand scanning process. For a subtractive color reproductionsystem37 with transparent (nonscattering) CMYK colorants,the reflectance of the CMYK halftone overlay can bewritten as

r (x, y; λ) = rW (λ)∏

i∈{C,M,Y,K }[ti (x, y; λ)]2

= rW (λ)∏

i∈{C,M,Y,K }[ti (λ)]2I h

i (x,y), (5)

where ti (λ) is the transmittance of the i’th colorant, andrW (λ) is the reflectance of the paper substrate at wavelengthλ. Note that the incoming light is attenuated by traversingeach colorant layer twice and also reflected by the paper sub-strate. The colorant transmittances are further modeled asper the Beer-Bouguer law (Ref. 38, Chap. 7) or the transpar-ent formulation of Kubelka-Munk theory (Ref. 38, Chap.7) as [ti (λ)]2 = 10−di (λ), where di (λ) is the optical den-sity for the i’th colorant layer (under to and fro passage).Equation (5) becomes

r (x, y; λ) = rW (λ)∏

i∈{C,M,Y,K }10−di (λ)I h

i (x,y)

= rW (λ)10−∑i∈{C,M,Y,K } di (λ)I h

i (x,y). (6)

The k’th color channel of the scanner I sk (x, y),

k ∈ {R, G, B}, measures the fraction of the illuminant lightreflected from the print within the transmittance band of thefilter k. Using a narrow spectral band assumption for thescanner spectral sensitivities, we approximate the recordedk th scanner channel image as

I sk (x, y) = I s

k (W )10− ∑i∈{C,M,Y,K } di (λk )I h

i (x,y), (7)

where the k’th scanner channel sensitivity is approximatedby an impulse (Dirac delta function) at wavelength λk , andI sk (W ) = rW (λk) is the scanned image value corresponding




to the paper substrate in the kth color channel. The scannedimage values can be converted to density as

d (x, y; λk) = dk(x, y) = − log10

[I sk (x, y)

I sk (W )

]

=∑

i∈{C,M,Y,K }di

k I hi (x, y), (8)

where dik = di (λk) is the optical density of the i’th colorant

layer in the k’th scanner channel. Note that the densityspace representation for the scanned images in Eq. (8)avoids interseparation moire since the colorants combineadditively in the density space. On the other hand, moirewould be observed in the reflectance space representationgiven in Eq. (5) as the colorants combine multiplicativelyin the reflectance space. Therefore, interseparation moirewould remain present in typical scanned clustered-dot colorhalftone images. Even though we recognize that the modelis an extremely idealized approximation, use of the modelprovides an improvement over an alternate ad hoc approachas we shall subsequently see.

The constituent contone images Ii (x, y), i ∈ {C, M,Y, K }, are band-limited to frequencies significantly smallerthan the corresponding fundamental screen frequencies.Hence, most of the energy carried by the individual halftonesis concentrated in narrow bands around the correspondingfundamental screen frequencies and lower order harmonics.As the colorants combine additively in the optical densityspace as per Eq. (8), the intensity of any frequency compo-nent in the Fourier spectrum of the overlay can be easily ob-tained by a weighted linear combination of the correspondingintensities in the Fourier spectra of the constituent halftones,where the weights correspond to the optical densities of thecolorants in the scanner channel under consideration. There-fore, the frequency components that carry most of the energyin the individual halftone separations continue to carry (de-pending on the optical density) a significant portion of theenergy in the colorant overlays as well. The interference re-sulting from the unwanted halftone separations can be elim-inated/reduced by removing these frequency components.For this purpose, we utilize narrow-band band-reject filtersHi (u, v) to eliminate the unwanted frequency components ofthe i’th separation, where u and v represent the frequencycoordinates along the horizontal and vertical directions, re-spectively. These filters may be efficiently implemented byutilizing the 2-D fast fourier transform (FFT). Our experi-mental results indicate that the interference resulting fromthe unwanted halftone separations is significantly reducedby eliminating the unwanted frequency components in asmall circular neighborhood [typically around 3 to 4 lines perinch (lpi) radius] of the corresponding fundamental screenfrequencies, their second-order harmonics, and their linearcombinations.

Note that in the already described filtering process, wedo not want the band-reject filter for the screen frequenciesof the i’th separation to eliminate the frequency componentsthat carry significant portions of energy in the other sep-arations. The band-reject filters described in the precedingparagraph eliminate only the lower order harmonics of theunwanted fundamental screen frequencies. Therefore, unlessthe frequency components of the desired halftone separation

intersect with the frequencies eliminated by the band-rejectfilters, these frequencies will be preserved. Note that for 2-D lattices defined in an underlying regular digital grid ofpixels, the least common frequency for the separations inconsideration is the least common right multiple39 of thebasis matrices Wi ’s. For rotated halftone orientations, thisfrequency generally appears at higher order harmonics ofthe fundamental screen frequencies. Thus, the elimination oflower order harmonics of the halftone corresponding to a col-orant does not cause the elimination of any lower order har-monics for the other colorant halftones. Note that this argu-ment and our proposed method are restricted to clustered-dotmutually rotated colorant separations. The argument and theproposed method based on using the differences in spatial pe-riodicities are not applicable, for instance, for dot-on/off-dotorientations.27 For dot-on/off-dot configurations, the halftoneseparations use the same fundamental screen frequency andspatial filtering to eliminate one colorant halftone will alsoresult in the removal of desired frequency components fromother colorant halftones as well.

In the following section, we describe the proposed CMYKhalftone separation estimation methodology using the densi-ties of the RGB color channels of the scanned images dk (x, y),k ∈ {R, G, B}. As we proceed, we also provide 1-D illustra-tive examples that help explain the proposed technique. Fora 1-D contone image Ii (x), i ∈ {C, M, Y, K }, the halftonespots are rectangular pulses, and for image values normalizedbetween 0 and 1 representing white and black, respectively,the halftone spot for the gray-level Ii (x) is represented assi [x ; Ii (x)] = rect[x/Ii (x)]. Note that the rectangular spotsresult from the halftoning process—a conversion from col-orant amount to colorant surface coverage in our idealizedmodel. The coefficient in Eq. (4) for this spot functioncan be computed as cIi (x)

n = Ii (x) sinc [nIi (x)] and Eq. (3)becomes35

I hi (x) = Ii (x)

{1 + 2

∞∑n=1

sinc [nIi (x)] cos [2π fi nx]

}, (9)

where fi is the fundamental screen frequency of the i’thseparation. Using this representation, we generate examplesof CMYK halftone separations shown in Fig. 8. For ourexample, we use the frequencies fC = 1/3, fM = 1/4,fY = 1/7, and fK = 1/5 (in arbitrary units). The estima-tion process works on the RGB scanner densities shown inFig. 9, where the optical densities of the CMYK colorantswere obtained using the scanned RGB image values of thecorresponding colorants measured for the printer used in ourexperiments. Thus, these do reflect the nonidealities in theprinting and scanning process that arise from the colorantinteractions and are not simply based on our extremely ide-alized model of Eq. (8).

3.2 Estimation of K Halftone SeparationAll scanner channels are jointly used for the estimation ofthe K halftone separation as K colorant absorbs nearly uni-formly across the three scanner RGB channels and thereforecontributes similarly to dk(x, y), k ∈ {R, G, B}. Apart fromthis use of spectral characteristics, the black (K) halftone sep-aration is estimated by exploiting primarily the spatial char-acteristics of the halftone separations. The overall estimationprocess is illustrated in Fig. 10. First, we add the scanner




0

1

x

Ih C(x

)

0

1

x

Ih M

(x)

0

1

x

Ih Y(x

)

0

1

x

Ih K

(x)

(a) (b)

(d)(c)

Fig. 8 Sample 1-D halftone patterns for the CMYK halftone separations. (a) Cyan; (b) magenta; (c) yellow; and (d) black.

channel densities together in dK (x, y) = ∑k∈R,G,B dk(x, y)

to obtain a monochrome representation for the scannedimage.

There are 24 = 16 possible overlapping combinations ofthe CMYK colorants on the printed paper that are referred toas Neugebauer primaries.37 As the Neugebauer primaries thatinclude the K colorant are expected to have larger densitythan the others, one might consider obtaining the estimateof the K halftone from dK (x, y) by a simple thresholdingoperation. The accuracy of this approach depends on the op-tical densities of the individual colorants. Among the eightNeugebauer primaries that include the black (K) colorant,the primary with the lowest optical density is the K colorantprinted alone. Among the remaining eight Neugebauer pri-maries that do not include the K colorant, the overlay ofC, M , and Y colorants has the highest optical density. There-fore, to not estimate the overlay of CMY colorants as Kcolorant, the constraint∑k∈{R,G,B}

dC MYk <

∑k∈{R,G,B}

d Kk (10)

Scan channel densities

<HC HYHM

dR(x, y)dK(x, y)

IhK(x, y)τK0

1

dG(x, y)

dB(x, y)

dK(x, y)

>

Fig. 10 Detailed overview of K halftone separation estimation ex-ploiting the differences in spatial periodicity.

0

1

2

3

x

dR(x

)

0

1

2

3

x

dG

(x)

0

1

2

3

x

dB

(x)

(a) (b)

(c)

Fig. 9 Predicted RGB scanner channel densities for the halftone images shown in Fig. 8. (a) Red; (b) green; and (c) blue.




u (lpi)

v(lpi

)(fC , φC

)↘

↙(fM , φM

)↙

(fY , φY

)

↑(fK , φK

)

−100 −50 0 50 100

100

50

0

−50

100

(a)

u (lpi)

v(lpi

)

−100 −50 0 50 100

100

50

0

−50

100

(b)

Fig. 11 Spatial filtering for estimation of K halftone IhK (x): (a) enlarged view of the log-magnitude Fourier

spectrum of dK (x, y) for the scanned “Hats” print used in our experiments and (b) the elimination ofunwanted frequency components of C, M, and Y halftone separations using the narrow-band band-reject filters shown in Figs. 24(a) to 24(c) in Sec. 5, respectively.

must be satisfied, where dC MYk denotes the density of

the overlay of CMY colorants in the k’th color channel ofthe scanned image. In most color printing systems, theseidentities differ by small margins, and estimating the Kcolorant purely based on spectral differences results in poorperformance, as shown in Sec. 3.4, for the printer used in ourexperiments.

To reduce the interference caused by CMY colorants, wenext use the band-reject filters Hi (u, v), i ∈ {C, M, Y }, toremove the unwanted frequency components of C, M , and Yhalftone separations from dK (x, y) as

dK (x, y) = F−1

⎧⎨⎩F{dK (x, y)}

∏i∈{C,M,Y }

Hi (u, v)

⎫⎬⎭ , (11)

where F denotes the (spatial) Fourier transform operation.Black (K ) halftone separation estimate is finally obtained by

I hK (x, y) =

{1 if dK (x, y) > τK

0 otherwise,(12)

where τK is a threshold value, which can be taken as theaverage of the minimum and maximum values in dK (x, y).

Figure 11 illustrates the elimination of the unwantedC, M , and Y halftone structures from the Fourier spectrumof the experimental data using the band-reject filters HC ,HM , and HY shown in Fig. 25 in Sec. 5. The K halftoneseparation estimate obtained using this method is shown inFig. 12. Note that the densities dK (x, y) and dK (x, y) aretransformed back to the reflectance space as I s

K (x, y) andI s

K (x, y) shown in Figs. 12(b) and 12(c), respectively, fora more meaningful representation. In addition, the estima-tion of K halftone separation using the 1-D densities shownin Fig. 9 is illustrated in Fig. 13. The elimination of thefundamental screen frequencies and the second-orderharmonics of the CMY separations is reflected in the

(a) (b) (c) (d)

Fig. 12 Sample estimation result for the K halftone separation for the region shown in Fig. 4(c) fromthe “Hats” halftone image used in our experiments: (a) the digital K halftone separation Ih

K (x, y),(b) the monochrome representation Is

K (x, y), (c) the filtered image IsK (x, y), and (d) the K halftone

estimate IhK (x, y).




0

2

4

6

8

x

dK

(x)

0

2

4

6

8

x

dK

(x)

0

1

xIh K

(x)

(a) (b)

(c)

Fig. 13 Illustration of the K halftone separation estimation using the scanner channel densities shownin Fig. 9: (a) the density for the monochrome representation dK = ∑

k∈{R,G,B} dk ,(b) the filtered density

d K , and (c) the K halftone estimate, which matches the original shown in Fig. 8(d).

corresponding representations given in Eq. (9) by changingthe lower bound of the summations to n = 3.

3.3 Estimation of C, M, and Y Halftone SeparationsThe efficacy of using the differences in spatial periodicity1

and colorant spectral differences28 for the estimation ofCMY halftone separations from the RGB channels of thescanned images has been shown previously for three-colorant CMY printing. For CMYK printing, however, theblack (K ) colorant introduces additional challenges so thatthese methods alone are insufficient to provide accurateestimates.

If the additive in density model of Eq. (8) were exact, thespatial filtering procedure to eliminate undesired halftoneseparations would provide accurate estimates of the C, M ,and Y halftones. However, the actual colorant interactionsdeviate significantly from the ideal model particularly, athigh densities associated with overlaps of the K colorant.As a result, when estimating, for example, the C halftone,removal of the frequency components of the K halftone sepa-ration from the RGB scans by spatial filtering also eliminatesparts of the C halftone that are overlaid with the K col-orant halftone. Thus, spatial filtering alone does not enablerecovery of the C, M , and Y halftones. An attempt to sepa-rate the halftones based purely on the spectral characteristics(i.e., scanner channel densities) faces similar challenges

because the scanned the image values for the Neugebauerprimaries including K differ only by small margins. We,therefore, address this problem by utilizing both techniquesjointly, where in addition to the scan RGB channel im-ages, we use the K halftone separation estimate for the es-timation of the complementary CMY halftone separations,respectively.

We describe only the estimation of the M halftone sep-aration, with examples illustrating the images obtained dur-ing the process. Other halftone separations can be estimatedsimilarly. A detailed overview of the M halftone separationestimation is illustrated in Fig. 14.

The M halftone separation is estimated from its comple-mentary G channel of the scanned image and the K halftoneseparation estimate obtained previously. First, an estimate ofthe overlap of M and K halftone separations is obtained byusing the G scan channel density dG(x, y). As the colorantscombine additively in the density space, the densities of theNeugebauer primaries that include both M and K colorants islarger than the densities of the constituent colorants. In otherwords, the overlays that include both M and K colorants willappear darker than the constituent colorants in the G colorchannel of the scanned image. Using this property, we obtainan estimate of the M K halftone overlap as

I hM K (x, y) =

{1 if dG(x, y) > τM K

0 otherwise,(13)

Halftone Invert

<

<>0

1τMK

+

HC HY

+1

IhMK(x, y)

dMK(x, y)G scan channel density

IhM(x, y)τM0

1

dM(x, y)

IhK(x, y)

−

IhK (x, y)

dG(x, y)

−

>

Fig. 14 Detailed overview of M separation estimation using the scanner G channel image and the Kseparation estimate.




M Y K CM CY CK MY MK YK CMY CMK CYK MYKCMYK W0

64

128

192

255

Neugebauer Primaries

Red Green Blue

C M Y K CM CY CK MY MK YK CMY CMK CYK MYKCMYK W0

0.5

1

1.5

2

2.5

Neugebauer Primaries

Red Green Blue

(a) (b)

Fig. 15 Scan RGB (a) reflectance space and (b) density space values for the 16 Neugebauer primariesfor the printer used in our experiments.

where τM K is a threshold value and, in our convention, thevalues 1 and 0 correspond, respectively, whether both Mand K toners/inks are, or are not, estimated to be depositedat the pixel position (x, y). The threshold τM K is chosenmidway between the G scanner channel densities for theNeugebauer primaries formed by the M K overlay and thatof the M or K (alone) primary, whichever is higher. Thesedensities can be measured from scans by segmenting andaveraging interior regions corresponding to the Neugebauerprimaries.

The accuracy of this estimation step depends on theoptical densities of the individual colorants in the scan Gchannel. There are four Neugebauer primaries that includeboth the magenta (M) and black (K ) colorants. In this group,the primary with the lowest optical density corresponds

to the overlay of M and K colorants alone. Among theremaining 12 Neugebauer primaries, the overlay of C, M ,and Y, and the overlay of C, Y , and K colorants have thehighest optical densities. To ensure that these primaries arenot incorrectly assigned as the M K Neugebauer primary, theconstraints

dC MYG < d M K

G and

dCY KG < d M K

G

(14)

must be satisfied, where dC MYG , dCY K

G , and d M KG are the G

channel densities for the Neugebauer primaries correspond-ing to the overlay of the CMY, CYK, and M K colorants,respectively. In addition, to be able to distinguish the opticaldensity of the overlap of M and K colorants from the optical

u (lpi)

v(lpi

)

(fC , φC

)↘

↙(fM , φM

)↙(fY , φY

)

↑(fK , φK

)

−100 −50 0 50 100

100

50

0

−50

100

(a)

u (lpi)

v(lpi

)

−100 −50 0 50 100

100

50

0

−50

100

(b)

Fig. 16 Spatial filtering for estimation of the density of M and K halftone overlay dM K (x, y): (a) en-larged view of the log-magnitude Fourier spectrum of the G scan channel density of the scanned“Hats” halftone print used in our experiments and (b) the elimination of unwanted frequency compo-nents of C and Y halftone separations using the band-reject filters shown in Figs. 4(a) and 24(c),respectively.




(a) (b) (c)

(d) (e) (f)

Fig. 17 Images obtained during the estimation of M halftone separation corresponding to the regionshown in Fig. 4(d) from the “Hats” halftone image used in our experiments: (a) the estimate of M andK colorant overlap Ih

M K (x, y), (b) the K halftone estimate obtained previously IhK (x, y), (c) the estimate

of K halftone structures that are not overlapping with M halftone structures IhK ′ (x, y), (d) the estimate

of M and K halftone overlay IsM K (x, y), (e) the image after removing the K only halftone structures by

IhK ′ masking Is

M (x, y), and (f) the M halftone estimate IhM K (x, y).

densities of both M and K colorants, neither of the opticaldensities of the M and K colorants in scan G channel can be∞, in other words, neither of the M and K colorants can haveperfect absorption, which typically is the case in most colorprinting systems. This can be observed in the experimentaldata in Fig. 15(b) that shows the scan RGB values forthe 16 Neugebauer primaries for the printer used in ourexperiments.

Next, the estimate I hM K (x, y) is subtracted from the pre-

viously obtained K halftone separation estimate I hK (x, y)

to get an estimate of I hK ′ (x, y) that represents K halftone

structures that are not overlapping with M halftone struc-tures. The unwanted frequency components of C and Yhalftone separations are then eliminated from the scan Gchannel density using the band-reject filters HC and HY ,respectively, by

dM K (x, y) = F−1

⎧⎨⎩F{dG(x, y)}

∏i∈{C,Y }

Hi (u, v)

⎫⎬⎭ . (15)

This image approximates the density of the overlay M andK halftone separations in the scanner G channel. Figure 16illustrates the elimination of the unwanted frequency compo-

nents of C and Y halftone structures in the Fourier spectrumof the density of the G channel of the experimental data us-ing the band-reject filters HC and HY shown in Sec. 5 inFigs. 24(a) and 24(c), respectively.

We then use I hK ′ (x, y) as a mask to clean out the K halftone

structures that do not overlap with M halftone structuresfrom dM K (x, y) as dM (x, y) = dM K [1 − I h

K ′ (x, y)]. The Mhalftone separation estimate is finally obtained by

I hM (x, y) =

{1 if dM (x, y) > τM

0 otherwise,(16)

where τM is a threshold that is set midway between the scan-ner G channel densities for the M (alone) Neugebauer pri-mary and the CY Neugebauer primary, which has the nexthigher density.

We previously showed in Fig. 4(e) a sample estimate forthe M halftone separation using the G color channel of thescanned image shown in Fig. 4(d) and the K halftone estimateshown in Fig. 12(d). To illustrate the method, the images ob-tained in intermediate stages of the proposed estimation pro-cess are shown in Fig. 17. Note that the densities dM K (x, y)and dM (x, y) are transformed back to the reflectance spaceas I s

M K (x, y) and I sM (x, y) shown in Figs. 17(d) and (e),




0

1

x

Ih M

,K(x

)

0

1

x

Ih K

′(x)

0

0.5

1

1.5

2

2.5

x

dM

(x)

0

1

x

Ih M

(x)

(a) (b)

(c) (d)

Fig. 18 Illustration of the M halftone separation estimation using the density shown for the G scanchannel in Fig. 9(b) and the K halftone estimate shown in Fig. 13(c): (a) the estimate of M K halftoneoverlap, (b) the estimate of K halftone separation not overlapping with M halftone, (c) the filtered densitydM K multiplied by [1 − Ih

K ′ (x)], and (d) the M halftone estimate, which matches the original shown inFig. 8(b).

respectively, for a more meaningful representation. In addi-tion, the estimation of M halftone separation using the 1-Dscan G channel density shown in Fig. 9(b) and the previ-ously obtained K halftone estimate shown in Fig. 13(c) isillustrated in Fig. 18. The elimination of the fundamentalscreen frequencies and the second-order harmonics of the Cand Y separations is reflected in the corresponding represen-tations given in Eq. (9) by changing the lower bound of thesummations to n = 3.

3.4 Performance Analysis: Accuracy of theEstimated Halftones

Given the dot gain and the geometric distortion in theprinting and other nonidealities in the print-scan process,it is difficult to experimentally evaluate the performance ofthe estimation process quantitatively. Therefore, we utilizea simulation framework, where the effect of print-scan issimulated by the Neugebauer model37 and spatial blur.The overview of the simulation is shown in Fig. 19.First, the individual colorant separations of the contoneimage IC,M,Y,K (x, y) are halftoned. Next, the effect of theprint-scan is simulated on the digital halftone I h

C,M,Y,K (x, y).Estimates of the halftone separations I h

C,M,Y,K (x, y) arethen obtained from I s

R,G,B(x, y) using our methodology, andfinally, the performance metrics for the estimates of eachseparation are calculated.

Twenty-four images from the Kodak PhotoCD image set,as shown in Fig. 20, were used in our simulations. TheCMYK versions of these images were obtained using Adobe R©

Photoshop R© CS4 with default CMYK settings. We thengenerated clustered-dot halftones for these images in 3.84-×2.56-in. format for a 600 dots per inch (dpi) CMYK colorprinter. Digital rotated clustered-dot CMYK halftone screenswith angular orientations close to the conventional singu-lar moire-free 15-, 75-, and 45-deg rotated clustered-dothalftones were utilized for our simulations.40 The funda-mental frequency vectors for these screens are shown in Fig.21. Since the objective was to evaluate the accuracy of theestimated halftones, no watermarks were utilized in thesesimulations.

The effect of print-scan is simulated for a 600 dpi CMYKcolor printer and RGB color scanner combination. For thispurpose, we utilize a Neugebauer model37 in scanner RGBspace. Binary CMYK printer values are mapped to contoneRGB scanner values through a look-up table (LUT) trans-formation. For this purpose, the average RGB values of 16Neugebauer primaries for the printer used in our experimentswere measured from 0.5-×0.5-in.2 printed patches and storedin the LUT. Figure 15(a) shows the measured scan RGBvalues for the 16 Neugebauer primaries. The CMYK valuesof each pixel in the color halftone are mapped to the RGBscanner values using the LUT. Note that since we use ac-tual measured values, spectral nonidealities in the colorant

IC,M,Y,K(x,y )

Performance

Print & Scan

Simulation Estimation

Halftone Separation

Accuracy Sensitivity Specificity

Halftoning

IsR,G,B(x,y )IhC,M,Y,K(x,y )

IhC,M,Y,K(x,y )

Evaluation

Fig. 19 Simulation model for evaluating the performance of the estimation process.




Fig. 20 Kodak PhotoCD image set used in the simulations. These images are referred to as Image 1to Image 24, enumerated from left-to-right and top-to-bottom, respectively.

interactions and in the scanner response are automaticallyincorporated in the modeling process. To also model spatialnonidealities, we performed a spatial blur on the individualRGB channels of the resulting image, analogous to themethodology of Ref. 41. For this purpose, a 7×7 Gaussianlow-pass filter (DC response normalized to 1) with standarddeviation approximately equal to 0.96 pixels was employed.For the 600-dpi printer resolution we assumed, this corre-sponds to a standard deviation of 40 μm, and the blur causesspreading in an approximately 120 μm radius. Note that amore severe blur increases the impact of the spatial interac-tions, and consequently, the accuracy of the estimates woulddecrease. Similarly, a milder blur brings the simulation closerto aspatiallyideal hard-dot setting,42, 43 and thus, the accuracyof the estimates would be higher.

Using the described methodology, we obtained estimatesof the individual halftone separations. The band-reject filtersdesigned for these frequency vectors are shown in Fig. 22. Forpurpose of visual comparison, enlarged views of the originalCMYK halftone separations and their estimates from a regionin Image 3 in Fig. 20 (also referred to as the “Hats” image inthe remaining parts of this paper) are shown in Fig. 23.

The performance of the described halftone separation esti-mation methodology is evaluated by the accuracy, sensitivity,and specificity, which are common performance measures forbinary data classification techniques.44 In our context, theyrepresent the percentage of overall, majority, and minoritypixels that are correctly classified, respectively, and are com-puted for each colorant as

accuracy = Ntp + Ntn

Ntp + Ntn + Nfp + Nfn×100,

sensitivity = Ntp

Ntp + Nfn×100,

(17)

specificity = Ntn

Ntn + Nfp×100,

where Ntp is the number of true positive predictions cor-responding to number of majority pixels that are predictedcorrectly, Ntn is the number of true negative predictions cor-responding to number of minority pixels that are predictedcorrectly, Nfp is the number of false positive predictions cor-responding to the number of minority pixels that are pre-dicted as majority, and Nfn is the number of false negativepredictions corresponding to the number of majority pixelsthat are predicted as minority. The estimation performancefor each colorant is listed in Table 1. Around 95% accuracyis achieved for C , around 98% accuracy is achieved forK halftone separations, and relatively lower accuracies areachieved for the M and Y halftone separations. This differ-ence is caused by the varying levels of spectral interferenceobserved in the scanner RGB channels. As we can see inFig. 2, relatively less interference is observed in the scannerR channel compared to the G and Y channels. Therefore,the C and K halftone separations can be estimated betterthan the M and Y halftone separations. Nevertheless, as wedemonstrate later in Sec. 5, these estimates enable successfuldetection of the embedded watermark patterns.

C: 47.4 lpi@ 18.4◦

K: 42.4 lpi@ 45◦

Y: 60 lpi@ 0◦

v

u

M: 47.4 lpi@ 71.6◦

Fig. 21 Frequency vectors for singular moire-free C, M, Y , and Khalftone screens.




u (lpi)

v(lpi

)

−300 0 300

300

0

−300

(a) HC (u, v)

u (lpi)

v(lpi

)

−300 0 300

300

0

−300

(b) HM (u, v)

u (lpi)

v(lpi

)

−300 0 300

300

0

−300

(c) HY (u, v)

Fig. 22 Narrow-band band-reject filters for eliminating the frequency components of the CMY halftoneseparations from the densities of the RGB color channels of the scanned images for the frequencyvector configuration shown in Fig. 21. (a) cyan, (b) magenta, and (c) yellow.

We compared our methodology against three other estima-tion methods. The first estimation method is based purely oncolorant spectral differences. All scanner channels are jointlyused for the estimation of CMYK halftone separations. Fora CMYK printed and RGB scanned image, the histogram inthe 3-D RGB space is expected to have 16 distinct pointscorresponding to the 16 Neugebauer primaries. Due to theprint-scan nonidealities, however, the histogram typically ex-hibits scattered points that form clusters around each of theNeugebauer primaries. With this motivation, we employeda vector-quantization-based estimation technique45, 46 to ob-tain the constituent halftone separations. For each pixel inthe print-scan simulated halftone image, the pixel is classi-fied to the Neugebauer primary to which the pixel valuesare closest in terms of Euclidean distance in RGB coordi-nates. Individual halftone separation estimates were obtainedby transforming the Neugebauer primary indices to corre-

sponding binary halftone separations. The RGB values ofthe Neugebauer primaries for the printer used in our experi-ments were previously shown in Fig. 15(a). The performanceof this methodology for the image set in Fig. 20 is listed inTable 2. It is readily seen that the performance of our pro-posed methodology is better than the VQ-based estimationtechnique. Specifically, for the C colorant, our method pro-vides significantly better estimates. For this colorant we seethat the VQ algorithm misclassifies some of the pixels thatinclude the C colorant as K and the ones that include theY colorant as the CMY overlay for the data from the printerused in our experiments.

The second method is based purely on spatial filtering.The estimation process of the K halftone separation is thesame as our method. For the estimation of the C, M , and Yhalftone separations, the unwanted frequency components ofK halftone separation are also eliminated by a corresponding

(a) IhC(x, y) (b) IhM (x, y) (c) IhY (x, y) (d) IhK(x, y)

(e) IhC(x, y) (f) IhM (x, y) (g) IhM (x, y) (h) IhK(x, y)

Fig. 23 Enlarged view of the digital halftone separations: (a) to (d) from a region in the “Hats” imageand (e) to (h) the corresponding estimates obtained using the differences in spatial periodicity andcolorant spectra. (a) and (e), cyan; (b) and (f), magenta; (c) and (g), yellow; (d) and (h), black.




Table 1 Performance of the proposed halftone separation estimation methodology for the image set shown in Fig. 20.

Accuracy (%) Sensitivity (%) Specificity (%)

Image No. C M Y K C M Y K C M Y K

1 95.5 89.6 90.4 97.8 97.4 87.8 97.0 99.4 93.3 91.6 78.1 97.8

2 95.6 94.1 96.5 99.1 99.3 96.4 99.2 99.5 86.6 83.0 54.8 99.1

3 95.4 88.8 90.6 97.5 93.3 92.4 97.4 99.1 97.5 85.2 78.0 97.5

4 95.7 91.4 89.7 98.1 98.1 95.8 97.8 99.6 91.8 82.9 74.8 98.1

5 93.2 85.6 86.1 98.1 89.3 91.4 94.7 98.9 97.8 78.1 70.6 98.1

6 96.4 90.1 91.4 98.1 98.2 90.8 96.5 99.5 94.3 88.8 84.1 98.1

7 95.2 88.4 90.4 97.0 93.8 86.2 97.3 99.2 96.5 91.0 75.8 97.0

8 95.3 90.0 88.3 97.9 97.9 88.7 82.9 99.3 92.3 91.5 94.6 97.9

9 97.0 92.9 91.5 96.8 96.9 93.7 88.8 99.4 97.1 91.5 94.8 96.8

10 96.6 91.4 89.5 97.1 96.1 90.9 84.9 99.2 97.2 92.2 94.5 97.1

11 94.7 87.6 87.9 96.9 93.2 83.7 95.7 99.0 96.6 91.6 75.5 96.9

12 98.5 94.9 96.5 99.3 99.2 97.2 95.3 99.9 97.3 88.1 97.6 99.3

13 94.6 86.4 90.8 98.1 91.4 85.1 96.6 99.2 97.7 88.0 78.1 98.1

14 94.4 86.6 88.4 98.1 91.4 92.0 96.1 99.3 97.6 80.9 71.7 98.1

15 94.7 89.7 88.5 98.5 98.6 85.1 94.6 99.3 89.3 94.3 81.1 98.5

16 94.9 87.3 88.1 97.3 93.4 83.7 95.3 99.1 96.9 91.6 78.7 97.3

17 93.5 84.5 85.0 97.8 91.2 93.4 93.9 99.2 96.7 73.8 71.7 97.8

18 92.4 82.6 87.7 98.0 88.5 91.5 96.5 98.7 97.7 71.0 62.0 98.0

19 95.7 88.8 91.3 97.3 97.5 87.6 96.1 99.2 93.8 90.4 84.8 97.3

20 96.9 92.1 93.9 99.1 99.3 92.9 93.4 99.7 89.9 89.8 94.9 99.1

21 96.4 90.9 91.7 97.8 95.1 90.0 87.9 99.5 97.9 91.9 95.6 97.8

22 95.9 88.6 91.3 98.3 98.0 86.9 96.8 99.5 93.6 90.7 80.9 98.3

23 96.0 89.2 91.8 98.1 98.3 87.2 97.1 99.4 93.4 91.4 78.3 98.1

24 94.7 87.2 87.9 98.1 92.0 83.8 96.0 99.4 97.4 91.3 76.6 98.1

band-reject filter HK (u, v) from the complementary R, G,and B scanner channels. The performance of this method-ology for the image set in Fig. 20 is listed in Table 3. Theestimates for the C, M, and Y colorants are generally lessaccurate than the ones obtained using the proposed joint ap-proach. This is due to the nonidealities observed in practice,as described in Sec. 3.3.

The last method was used in our preliminary work on thisproblem.2 It simply uses the methods described in Secs. 3.2and 3.3 with the distinction that the images are processed inthe reflectance space instead of the density space processingsuggested by our model. The accuracy for the reflectance-space-based estimation method and the improvement withthe density space processing is listed in Table 4. We cansee that processing the images in the optical density space

generally provides better estimates. Specifically, the accu-racy for the K halftone estimates is significantly higher withthe processing in density. The estimates for the M and Yhalftones show smaller improvements, whereas those for theC halftones show no major improvement.

4 Color Halftone Watermarking using CPMHaving described the overall framework in a general setting,we next demonstrate a specific and concrete instantiationbased on CPM for monochrome watermark embedding.21 Inthis specific instantiation, the watermarks wi can be thoughtof as visual patterns represented as low-bandwidth (in com-parison with the halftone frequency) images. Large-font text




Table 2 Performance of the vector-quantization (VQ)-based halftone separation estimation method for the image set shown in Fig. 20.



1 83.9 87.2 86.7 90.2 80.3 84.3 96.6 98.9 87.9 90.4 68.4 58.2

2 81.6 90.9 95.4 87.4 82.5 97.5 98.5 99.1 79.2 57.9 46.8 54.3

3 83.3 85.6 85.6 90.7 87.2 89.5 94.7 98.8 79.2 81.7 68.8 64.8

4 83.3 89.3 87.2 89.6 82.1 95.7 95.8 98.9 85.3 76.7 71.4 59.4

5 77.7 83.4 84.1 89.5 80.7 87.9 89.6 99.0 74.0 77.7 74.2 73.9

6 86.3 89.3 89.2 93.5 86.1 90.4 95.5 98.3 86.4 86.9 80.2 64.5

7 84.6 87.0 85.6 91.3 87.8 86.5 96.3 98.6 81.4 87.6 63.1 61.6

8 85.3 86.6 86.4 92.1 84.8 85.3 83.5 97.9 86.0 88.1 89.7 64.9

9 91.4 88.5 89.6 94.9 93.9 88.5 87.5 98.2 88.7 88.4 92.1 51.1

10 88.5 86.9 87.1 93.3 92.5 86.5 81.6 99.1 83.4 87.4 93.1 55.5

11 81.5 84.8 82.9 90.4 86.9 82.0 92.2 99.2 74.1 87.7 67.9 69.0

12 94.3 93.8 92.7 97.3 95.2 95.3 90.4 99.3 92.7 89.4 95.0 62.5

13 80.7 85.9 87.4 90.1 83.2 84.7 94.3 98.5 78.1 87.5 72.2 66.2

14 81.0 85.2 85.7 90.9 82.9 88.1 92.4 98.9 79.0 82.2 71.4 72.7

15 82.5 88.1 87.2 92.2 84.8 86.5 87.5 99.2 79.4 89.7 86.7 78.1

16 82.3 84.5 85.3 91.4 86.6 82.4 91.8 99.3 76.7 86.8 76.8 67.7

17 79.1 82.0 82.3 90.7 82.4 85.6 85.9 99.1 74.6 77.5 76.7 78.6

18 74.0 82.0 84.0 89.2 77.0 86.7 89.2 98.7 69.9 75.9 68.7 77.3

19 85.1 85.7 86.5 91.3 81.8 82.8 92.9 98.7 88.5 89.3 77.8 60.2

20 89.6 92.2 91.8 95.4 92.8 94.0 93.8 99.0 79.9 86.3 88.3 74.3

21 87.1 87.5 89.8 93.6 91.0 85.9 88.6 98.6 82.5 89.5 90.9 65.1

22 84.1 87.7 89.0 91.6 81.5 85.1 94.9 97.9 86.8 90.8 77.9 62.8

23 84.0 89.0 90.5 92.2 82.6 86.3 95.4 98.2 85.5 92.0 77.9 66.0

24 81.5 85.7 86.5 90.7 84.3 83.4 93.1 98.7 78.6 88.4 77.2 66.9

rendered as a bilevel image is one example of such visualpatterns that we utilize for our illustration.

Continuous phase-modulated halftoning21 alters thehalftone phase through a modulation of the phase of an under-lying analytic halftone threshold function. For watermarkingapplications, this method is similar to other phase-shift-basedclustered-dot halftone watermarking methods12, 15, 17 but hasthe advantage that the watermark pattern may be decidedon dynamically just prior to halftoning, as opposed to thealternative embedding strategies where the pattern must beincorporated in the design of the halftone thresholds, whichare therefore static. For completeness, we briefly summarizethe embedding process here in our specific context of colorhalftoning. The halftone for the i’th separation, I h

i (x, y) is

assumed to be obtained by comparing the contone imagevalues Ii (x, y) against a periodic halftone threshold functionTi (x, y) as

I hi (x, y) =

{1 if Ii (x, y) < Ti (x, y)0 otherwise.

(18)

To describe our watermark embedding technique, we usea modified version of the analytic description proposed byPellar and Green47 and Pellar.48 We assume that the fun-damental screen frequency for the i’th halftone separation,i ∈ {C, M, Y, K }, is fi lpi that is oriented along the angleφi . This description assumes orthogonal halftones, where




Table 3 Performance of the spatial-filtering-based halftone separation estimation method for the image set shown in Fig. 20.



1 87.7 86.7 87.6 97.8 87.6 82.7 97.8 99.4 87.8 91.0 68.8 92.3

2 80.5 92.5 96.7 99.1 93.2 98.4 99.8 99.5 49.3 63.4 49.1 97.9

3 84.6 81.5 85.0 97.5 94.3 90.4 96.2 99.1 74.3 72.4 64.2 93.0

4 80.7 86.0 87.4 98.1 75.2 93.5 98.0 99.5 89.7 71.5 68.0 93.9

5 81.3 80.4 84.9 98.0 96.0 94.6 98.2 98.8 63.7 62.4 61.1 96.8

6 89.8 82.9 88.7 98.1 85.2 81.3 98.4 99.5 95.0 86.0 74.8 91.6

7 86.7 83.2 86.7 97.0 93.8 76.9 98.0 99.2 79.5 90.8 62.9 89.4

8 88.8 85.9 84.8 97.9 82.5 80.0 75.0 99.3 96.1 93.2 96.1 93.1

9 92.3 89.8 87.1 96.8 96.3 88.0 80.6 99.5 88.1 92.5 95.1 73.6

10 90.4 84.9 83.1 97.1 94.3 82.8 73.2 99.2 85.7 88.0 93.9 85.2

11 84.4 81.3 82.5 96.9 94.8 73.0 94.5 98.9 70.3 89.7 63.3 92.3

12 93.7 94.9 94.8 99.3 92.0 94.3 92.6 99.9 96.7 96.6 97.0 91.5

13 84.5 80.4 90.2 98.0 94.4 72.4 99.1 99.2 74.6 91.2 71.0 95.2

14 83.6 81.0 86.6 98.0 95.5 93.7 98.3 99.2 71.2 67.8 61.5 95.6

15 83.0 86.1 87.6 98.5 73.6 76.4 98.0 99.3 95.9 95.9 75.1 97.1

16 86.2 80.1 83.0 97.3 94.9 73.2 93.6 99.1 75.0 88.2 69.2 92.1

17 81.0 78.2 83.0 97.7 95.7 93.9 96.8 99.1 61.0 59.1 62.3 95.9

18 79.5 75.4 88.0 97.9 96.0 94.5 99.5 98.7 57.3 50.8 54.3 97.1

19 86.7 82.8 85.4 97.3 79.0 76.0 92.2 99.2 94.8 91.1 76.4 90.4

20 91.3 88.9 93.9 99.1 90.2 88.4 91.0 99.7 94.6 90.5 98.7 96.4

21 89.1 87.8 90.1 97.8 96.2 83.3 82.4 99.5 80.6 93.7 98.1 90.4

22 85.2 80.7 90.5 98.3 77.9 74.4 98.6 99.5 93.3 88.2 75.4 94.1

23 86.2 81.9 91.4 98.1 78.8 75.4 99.0 99.5 94.4 89.1 72.1 93.6

24 84.6 79.7 85.3 98.0 93.7 72.3 97.6 99.3 75.3 88.5 68.3 94.7

the matrix of frequency vectors for each halftone separa-tion is composed of two orthogonal vectors of equal length.Note that more general descriptions are also possible.33 Then,assuming input continuous tone image values that are nor-malized to range between −1 and 1, the correspondingthreshold function for the i’th separation is

Ti (x, y) = cos

(2π

fi√2

x ′)

cos

(2π

fi√2

y′)

, (19)

where[x ′

y′

]=

[cos

(φi − π

4

) − sin(φi − π

4

)sin

(φi − π

4

)cos

(φi − π

4

)] [

x

y

]. (20)

A watermark pattern is embedded in the halftone bymodulating the phase of one of the cosine function argu-ments by incorporating a spatially varying phase function�i (x, y), which enables phase variations along the corre-sponding halftone orientation. The modified version of thehalftone threshold function is given by

Ti (x, y) = cos

[2π

fi√2

x ′ + �i (x, y)

]cos

(2π

fi√2

y′)

.

(21)

Discontinuities in the phase term �i usually cause un-desirable visible artifacts in the halftone separation. Thespatial continuity and smoothness of �i (x, y) is imposed




Table 4 Accuracy for the reflectance-space-based processing andimprovement in accuracy for the density-space-based estimationmethod over the reflectance-space-based estimation for the imageset shown in Fig. 20.

Accuracy (%) for Improvement with

Reflectance Space Density Space

Processing Processing

Image No. C M Y K C M Y K

1 95.4 88.4 88.5 95.2 0.1 1.2 1.9 2.6

2 96.5 93.8 95.6 95.0 − 0.9 0.3 0.9 4.1

3 95.4 88.0 88.3 94.3 0 0.8 2.3 3.2

4 96.1 91.0 88.8 95.6 − 0.4 0.4 0.9 2.5

5 92.7 84.5 84.4 93.7 0.5 1.1 1.7 4.4

6 96.7 90.0 90.3 95.8 − 0.3 0.1 1.1 2.3

7 95.5 88.1 88.9 95.0 − 0.3 0.3 1.5 2.0

8 95.3 89.4 87.5 94.8 0 0.6 0.8 3.1

9 97.2 92.7 91.7 95.6 − 0.2 0.2 − 0.2 1.2

10 97.1 91.4 89.9 95.6 − 0.5 0 -0.4 1.5

11 94.4 87.0 87.2 94.2 0.3 0.6 0.7 2.7

12 98.5 94.5 95.8 98.4 0 0.4 0.7 0.9

13 94.2 86.2 89.3 94.7 0.4 0.2 1.5 3.4

14 93.9 86.0 86.9 94.6 0.5 0.6 1.5 3.5

15 94.0 88.2 87.0 94.8 0.7 1.5 1.5 3.7

16 95.0 87.2 86.8 94.4 − 0.1 0.1 1.3 2.9

17 92.6 83.0 83.3 93.6 0.9 1.5 1.7 4.2

18 92.0 81.9 85.7 92.8 0.4 0.7 2.0 5.2

19 95.9 88.6 89.8 95.4 − 0.2 0.2 1.5 1.9

20 96.6 92.0 93.2 97.2 0.3 0.1 0.7 1.9

21 96.2 90.4 90.3 95.9 0.2 0.5 1.4 1.9

22 95.9 88.7 90.1 95.5 0 − 0.1 1.2 2.8

23 96.6 89.3 90.9 94.7 − 0.6 − 0.1 0.9 3.4

24 94.8 87.1 86.5 94.5 − 0.1 0.1 1.4 3.6

to minimize these artifacts. The technique is therefore re-ferred to as CPM halftoning. A visual watermark pat-tern such as a bilevel text image serves as the watermarkbased on which the phase modulation term �i for the i’thhalftone separation is determined. Typically, the bilevel wa-termark is smoothed using a spatial blur to ensure conti-nuity of the phase and the resulting values wi (x, y) arenormalized between (0, κ], κ ≤ π , to introduce a phase

K: 54.6 lpi@ 15◦

v

u

C: 45.8 lpi@ 82◦

M: 47 lpi@ 38.5◦

Y: 48.4 lpi@ 63◦

Fig. 24 Frequency vectors for stable moire-free C, M, Y , and Khalftone screens.

change of κ rad between the two levels of the watermarkedimage.

The digitally embedded watermark pattern in each sepa-ration can be retrieved by overlaying the halftone separationwith a corresponding “decoder mask,” which is a halftoneimage that has the same fundamental screen frequency as thewatermarked separation in consideration. It has been analyt-ically demonstrated that the visible pattern after this overlayresembles the embedded watermark pattern, which enablesvisual detection.21 Typically, the decoder mask has 50% areacoverage and κ is set to π to enable maximal contrast inthe detected watermark.21 In our proposed framework, anelectronic decoder mask corresponding to the 50% area cov-erage is used to extract the watermark patterns from theestimates of the individual colorant halftones. The overlayoperation is implemented in digital computation by multi-plying together “transmittances” of the decoder mask and theestimated halftone, each being set to 0 for colorant-coveredregions and 1 otherwise.

5 Experimental ResultsOur experimental setup utilized an electrophotographicCMYK color printer with an addressability of 600×600dpi.Stable moire-free screen frequency configurations are soughtto prevent self-detection with CPM (see the details in theappendix for a discussion on self-detection). For this pur-pose, we used the screen frequencies shown in Fig. 24,which are the scaled versions of the vectors described bySchoppmeyer.49 Note that alternatively the methodology de-veloped by Amidror et al.32 can be employed to determineadditional stable moire-free configurations.

In each of the C, M, Y , and K colorant separations,an independent watermark pattern was embedded viaCPM. The resulting color halftone was printed on ourtest printer and scanned with a flatbed RGB scanner withan optical resolution of 600×600 dpi. Estimates of theC, M, Y , and K halftones were obtained according to theproposed methodology using the R, G, and B channels ofthe scanned image. The band-reject filters designed for thefrequency vectors shown in Fig. 24 are shown in Fig. 25.The embedded watermark patterns are finally detected byoverlaying the estimates with the corresponding decodermasks.




u (lpi)

v(lpi

)

−300 0 300

300

0

−300

(a) HC (u, v)

u (lpi)

v(lpi

)

−300 0 300

300

0

−300

(b) HM (u, v)

u (lpi)

v(lpi

)

−300 0 300

300

0

−300

(c) HY (u, v)

Fig. 25 Narrow-band band-reject filters for eliminating the frequency components of the CMY halftoneseparations from the densities of the RGB color channels of the scanned images for the frequencyvector configuration shown in Fig. 24. (a) cyan, (b) magenta, and (c) yellow.

(a) wC (b) wM (c) wY (d) wK

Fig. 26 Watermark patterns for the C, M, Y , and K colorant separations. (a) cyan, (b) magenta, (c) yellow, and (d) black.

(a) (b) (c) (d)

(e) (f) (g) (h)

Fig. 27 Detection results on the CMYK halftone separation estimates obtained using our methodology:(a) to (d) the detection results for the “Hats” image (e) to (h) the detection results for the “Motorcycles”image. Artifacts due to rescreening in the printing process may not allow clear observation of thedetected watermark pattern in the printed version of these figures. Please refer to the electronic versionof this paper to clearly observe the detection results. (a) and (e), cyan; (b) and (f), magenta; (c) and (g),yellow; (d) and (h), black.




(a) (b) (c) (d)

(e) (f) (g) (h)

Fig. 28 Detection results on the initial contone scanned RGB channel images: (a) to (d) the detectionresults for the “Hats” image and (e) to (h) the detection results for the “Motorcycles” image. Artifactsdue to rescreening in the printing process may not allow clear observation of the detected watermarkpattern in the printed version of these figures. Please refer to the electronic version of this paper toclearly observe the detection results. (a) and (e), cyan; (b) and (f), magenta; (c) and (g), yellow; (d) and(h), black.

We illustrate the operation of the proposed methodol-ogy using the “Hats” and “Motorcycles” images that arereferred to as Image 3 and Image 5 in Fig. 20, respec-tively. The printed images were generated in 5.12-×3.41-in.format and the watermark patterns shown in Fig. 26 wereembedded in the corresponding halftone separations. Fig-ure 27 shows the detection results obtained for our exam-ple. The watermark patterns are generally visible through-out the images subject to variations due to the content ofthe corresponding contone image. Specifically, watermarksembedded in the highlights and shadows are harder to de-tect than the watermarks embedded in the midtones—as isexpected from the analysis of CPM watermark detection.21

This effect is particularly observed in the Y and K sepa-rations.The Y separation of the both images includes manydark regions with nearly 100% coverage, and the K channelof the “Hats” image includes many bright regions with nearly0% coverage, which do not allow for meaningful watermarkembedding and extraction due to the absence of halftonestructure.21

For purpose of comparison, we also performed detectionusing the initial contone scanner channel images that ex-hibit strong interseparation colorant interactions. Note thatthe demodulation provided by the electronic decoder maskautomatically forms a crude form of spatial filtering in theseimages because it selectively translates the frequencies cor-responding to the mask lattice to low frequencies,21 which infact correspond to the colorant separation of interest, whereasthis is not the case for other colorant separations. The water-mark patterns embedded in the C, M , and Y halftone sepa-rations were detected using the complementary R, G, and Bcolor channels of the scanned images, respectively, and thewatermark pattern in the K halftone separation was detectedusing the spectral average of the RGB color channels of thescanned images. Figure 28 shows the detection results. Wecan see that watermarks are visible in some regions wherethe spectral interference is low, though with rather low con-trast. Specifically, it is very hard to detect the watermark pat-

tern embedded in the K separation. Comparing Fig. 28 withFig. 27, we see that our proposed method enables detection ofthe embedded watermark patterns with much higher contrast.

6 ConclusionWe proposed a methodology for color halftone watermark-ing that embeds, via a monochrome halftone watermarkingmethod, watermarks in the halftone colorant separations of aclustered-dot color halftone. We demonstrated that by care-fully selecting halftone periodicities, the spatial character-istics of the separations and the spectral characteristics ofthe colorants can be used to estimate the individual CMYKhalftone separations from scanned RGB images, enabling ef-fective detection of the embedded watermarks despite theapparent degeneracy in estimating four colorant halftonesfrom three scanner channels and other undesired colorantinteractions. We demonstrated the efficacy of this method-ology using CPM for the embedding and detection of in-dependent per-separation watermarks—where the proposedmethod provides much higher watermark contrast at the de-tector than the alternative of direct detection. The proposedestimation technique can also be used as a general method forestimating CMYK halftones from RGB scans. For this pur-pose, we evaluated the accuracy of our estimates in a simula-tion framework and compared our results against alternativeestimation methods. Our evaluations indicate the proposedmethod provides satisfactory estimation accuracies for theindividual CMYK halftone separations and performs betterthan the alternatives.

Appendix: CPM Watermark Self-Detection forSingular Halftone ConfigurationsWhen the individual separations are subjected to phase mod-ulation for the purpose of watermark embedding, their instan-taneous frequencies in different spatial regions vary around anarrow region about their nominal values. For a monochromehalftone with a suitably designed CPM watermark, this fre-




Fig. 29 Self-detectable watermark patterns in a singular moire-freehalftone screen configuration. Artifacts due to rescreening in the print-ing process may not allow clear observation of the self-detected wa-termark patterns in the printed version of this figure. Please referto the electronic version of this paper to clearly observe the self-detectability. (Color online only.)

quency variation is imperceptible.21 The overlay of colorhalftone separations, however, typically exhibits moire atthe sum and difference frequencies of the constituent screenfrequencies.32 Therefore, when considering the halftones inoverlay, it is also necessary to ensure that the resulting moireis also imperceptible in the presence of the phase modulation.This requires a careful selection of the set of frequency vec-tors for the color halftone configuration. If an integer linearcombination of the constituent halftone frequency vectorsforms a closed polygon, the configuration is referred to assingular.32 For singular configurations, where the integer co-efficients resulting in singularity are small (for example, ±1or ±2), the phase modulation causes significant deviationsaround the dc frequency, visible as objectionable color shiftslocalized around the modulated regions.33, 50 These causean undesired effect that the embedded watermark patternsbecome visible (without any detection) in the printed colorhalftone image. We refer to this effect in the singular halftoneconfigurations as watermark self-detection.

Watermark self-detection is specifically observed in thecase for the conventional analog 15-, 75-, and 45-degequifrequency rotated clustered-dot and dot-on/off-dot colorhalftone geometries. To illustrate self-detection, we employthe set of frequency vectors shown in Fig. 21, which repre-sent a singular moire-free configuration.32 Using the CPMmethodology, we embedded the watermark patterns shownin Fig. 26 in the C, M, Y , and K separations of the Hats im-age. Figure 29 shows a downsampled version of the digitalCMYK halftone in which the watermark patterns are visiblewithout any detection.

With this motivation, for color applications of phase-modulation-based halftone embedding, it is preferable tochoose configurations that are stable, from a moire stand-point, against small variations in frequency, as opposed tosingular moire-free geometries.

AcknowledgmentThis work was supported in part by an award from the Xe-rox Foundation and by a grant from New York State Officeof Science, Technology & Academic Research (NYSTAR)through the Center for Electronic Imaging Systems (CEIS).

References1. B. Oztan and G. Sharma, “Clustered-dot color halftone watermarks,”

in Proc. IS&T/SID 16th Color Imaging Conf. Color Science and En-gineering: Systems, Technologies, Applications, pp. 99–104, Portland,OR (2008).

2. B. Oztan and G. Sharma, “Clustered-dot color halftone watermarksusing spatial frequency and color separability,” in Color Imaging XV:Processing, Hardcopy, and Applications, R. Eschbach and G. G. Marcu,Eds., Proc. SPIE 7528, 75280Y-1–12 (2010).

3. I. Cox, M. Miller, J. Bloom, J. Fridrich, and T. Kalker, DigitalWatermarking and Steganography, 2nd ed., Morgan Kaufmann, SanFrancisco (2007).

4. Special issue on digital watermarking, IEEE Signal Proc. Mag. 17(2000).

5. C. I. Podilchuk and E. J. Delp, “Digital watermarking: algorithms andapplications,” IEEE Signal Proc. Mag. 18, 33–46 (2001).

6. G. C. Langelaar, I. Setyawan, and R. L. Lagendijk, “Watermarkingdigital image and video data,” IEEE Signal Proc. Mag. 17, 20–46(2000).

7. R. L. van Renesse, Optical Document Security, 3rd ed., Artech House,Boston (2005).

8. I. Amidror, S. Chosson, and R. D. Hersch,“Moire methods for theprotection of documents and products: a short survey,” J. Phys. Conf.Ser. 77(1), 012001 (2007).

9. C. M. Briquet, Les Filigranes: Dictionaire Histoire Des Marques DuePapier Des Leur Appartion Vers 1282, Jusquen 1600, 2nd ed., Karl W.Hiersemann, Leipzig (1923).

10. K. T. Knox and S. Wang, “Digital watermarks using stochastic screens,”in Color Imaging: Device Independent Color, Color Hardcopy, andGraphic Arts II, G. B. Beretta and R. Eschbach, Eds., Proc. SPIE 3018,316–322 (1997).

11. Z. Baharav and D. Shaked, “Watermarking of dither halftoned images,”in Security and Watermarking of Multimedia Contents, P. W. Wong andE. J. Delp, Eds., Proc. SPIE 3657, 307–316 (1999).

12. S. Wang and K. T. Knox, “Embedding digital watermarks in halftonescreens,” in Security and Watermarking of Multimedia Contents II,P. W. Wong and E. J. Delp, Eds., Proc. SPIE 3971, 218–227 (2000).

13. H. Z. Hel-Or, “Watermarking and copyright labeling of printed im-ages,” J. Electron. Imaging 10, 794–803 (2001).

14. I. Amidror, “New print-based security strategy for the protection ofvaluable documents and products using moire intensity profiles,” inOptical Security and Counterfeit Deterrence Technique IV, R. L. vanRenesse, Ed., Proc. SPIE 4677, 89–100 (2002).

15. G. Sharma and S. Wang, “Show-through watermarking of duplexprinted documents,” in Security, Steganography, and Watermarkingof Multimedia Contents VI, E. J. Delp and P. W. Wong, Eds., Proc.SPIE 5306, 670–684 (2004).

16. C. Liu, S. Wang, and B. Xu, “Authenticate your digital prints withGlossmark images,” in Proc. IS&T NIP20: Int. Conf. on Digital Print-ing Technologies, pp. 312–316 (2004).

17. S. Huang and J. K. Wu, “Optical watermarking for printed documentauthentication,” IEEE Trans. Inf. Forens. Secur. 2, 164–173 (2007).

18. O. Bulan, V. Monga, G. Sharma, and B. Oztan, “Data embedding inhardcopy images via halftone-dot orientation modulation,” in Security,Forensics, Steganography, and Watermarking of Multimedia ContentsX, E. J. Delp, P. W. Wong, J. Dittmann, and N. D. Memon, Eds., Proc.SPIE 6819, 68190C-1–12 (2008).

19. J.-M. Guo, S.-C. Pei, and H. Lee, “Paired subimage matching water-marking method on ordered dither images and its high-quality progres-sive coding,” IEEE Trans. Multimed. 10, 16–30 (2008).

20. J.-M. Guo and Y.-F. Liu, “Hiding multitone watermarks in halftoneimages,” IEEE Multimed. 17, 34–43 (2010).

21. B. Oztan and G. Sharma, “Continuous phase-modulated halftones,”IEEE Trans. Image Process. 18, 2718–2734 (2009).

22. G. Sharma, R. P. Loce, S. J. Harrington, and Y. Zhang, “Illuminantmultiplexed imaging: special effects using GCR,” in Proc. IS&T/SID11th Color Imaging Conf. Color Science and Engineering: Sys-tems, Technologies, and Applications, pp. 266–271, Scottsdale, AZ(2003).

23. R. L. de Queiroz, K. M. Braun, and R. P. Loce, “Detecting spatiallyvarying gray component replacement with application in watermarkingprinted images,”J. Electron. Imaging 14, 033016,1–9 (2005).

24. N. Rudaz and R. D. Hersch, “Protecting identity documents by mi-crostructure color differences,” J. Electron. Imaging 13, 315–323(2004).

25. R. D. Hersch, P. Donze, and S. Chosson, “Color images visible underUV light,” ACM Trans. Graph. 26(3), 75.1–75.9 (2007).

26. R. Bala, R. Eschbach, and Y. Zhao, “Substrate fluorescence: bane orboon?” in Proc. IS&T/SID 15th Color Imaging Conf. Color Scienceand Engineering: Systems, Technologies, Applications, pp. 261–266,Albuquerque, NM (2007).

27. C. M. Hains, S. Wang, and K. T. Knox, “Digital color halftones,”Chap. 6. in Digital Color Imaging Handbook, G. Sharma, Ed., CRCPress, Boca Raton, FL (2003).



http://dx.doi.org/10.1109/79.939835

http://dx.doi.org/10.1109/79.879337

http://dx.doi.org/10.1088/1742-6596/77/1/012001

http://dx.doi.org/10.1088/1742-6596/77/1/012001

http://dx.doi.org/10.1117/1.1382612

http://dx.doi.org/10.1109/MMUL.2010.14

http://dx.doi.org/10.1109/TIP.2009.2028367

http://dx.doi.org/10.1117/1.2001671

http://dx.doi.org/10.1117/1.1687731

http://dx.doi.org/10.1145/1276377.1276471


28. O. Bulan, V. Monga, and G. Sharma, “High capacity color barcodesusing dot orientation and color separability,” in Media Forensics andSecurity XI, E. J. D. III, J. Dittmann, N. D. Memon, and P. W. Wong,Eds., Proc. SPIE 7254, 725417-1–7 (2009).

29. J. A. C. Yule, Principles of Color Reproduction, Applied to Photome-chanical Reproduction, Color Photography, and the Ink, Paper, andOther Related Industries, Wiley, New York (1967).

30. J. W. S. Cassels, An Introduction to the Geometry of Numbers, Springer,Berlin (1959).

31. A. Papoulis, Systems and Transforms with Applications in Optics,McGraw Hill, New York (1968).

32. I. Amidror, R. D. Hersch, and V. Ostromoukhov, “Spectral analysisand minimization of moire patterns in color separation,” J. Electron.Imaging 3, 295–317 (1994).

33. B. Oztan, G. Sharma, and R. P. Loce, “Misregistration sensitivity inclustered-dot color halftones,” J. Electron. Imaging 17, 023004,1–30(2008).

34. E. Dubois, “The sampling and reconstruction of time-varying im-agery with application in video systems,” Proc. IEEE 73, 502–522(1985).

35. D. Kermisch and P. G. Roetling, “Fourier spectrum of halftone images,”J. Opt. Soc. Am. 65, 716–723 (1975).

36. T. S. Rao, G. R. Arce, and J. P. Allebach, “Analysis of ordered ditherfor arbitrary sampling lattices and screen periodicities,” IEEE Trans.Acoust. Speech Signal Proc. 38, 1981–2000 (1990).

37. G. Sharma, “Color fundamentals for digital imaging,” Chap. 1 in Digi-tal Color Imaging Handbook, G. Sharma, Ed., CRC Press, Boca Raton,FL (2003).

38. F. Grum and C. J. Bartleson, Eds., Optical Radiation Measurements:Color Measurement, Vol. 2, Academic Press, New York (1980).

39. C. C. MacDuffee, The Theory of Matrices, Chelsea, New York(1946).

40. P. Delabastita, “Screening techniques, moire in four color printing,” inTAGA Proc., pp. 44–65 (1992).

41. S. Gustavson, “Color gamut of halftone reproduction,” J. Imaging Sci.Technol. 41, 283–290 (1997).

42. J. P. Allebach, “Binary display of images when spot size exceeds stepsize,” Appl. Opt. 19, 2513–2519 (1980).

43. T. N. Pappas and D. L. Neuhoff, “Printer models and error diffusion,”IEEE Trans. Image Process. 4, 66–80 (1995).

44. J. P. M. DeSa, Pattern Recognition: Concepts, Methods, and Applica-tions, Springer-Verlag, Berlin (2001).

45. P. Heckbert, “Color image quantization for frame buffer display,” Com-put. Graph. 16, 297–307 (1982).

46. L. Brun and A. Tremeau, “Color quantization,” Chap. 9 in DigitalColor Imaging Handbook, G. Sharma, Ed., CRC Press, Boca Raton,FL (2003).

47. R. J. Pellar and L. D. Green, “Electronic halftone generator,” U.S.Patent No. 4149183 (1979).

48. R. J. Pellar, “Electronic halftone generator,” U.S. Patent No. 4196451(1980).

49. J. Schoppmeyer, “Screen systems for multicolor printing,” U.S. PatentNo. 4537470 (1985).

50. I. Amidror and R. D. Hersch, “Neugebauer and Demichel: depen-dence and independence in n-screen superpositions for colour print-ing,” Color Res. Appl. 25, 267–277 (2000).

Basak Oztan received his BS degree (withhigh honors) in electrical and electron-ics engineering from Middle East Tech-nical University, Ankara, Turkey, in 2003and his MS and PhD degrees in elec-trical and computer engineering from theUniversity of Rochester, New York, in2004 and 2010, respectively. He is cur-rently a postdoctoral research associatewith the Department of Computer Science,Rensselaer Polytechnic Institute, Troy, New

York. He was a research intern with Xerox Webster Research Cen-ter, Webster, New York, during the summers of 2005 and 2006. He isthe recipient of a student paper award at 2006 IEEE International onConference on Acoustics, Speech, and Signal Processing (ICASSP)in the image and multidimensional signal processing category. Hisresearch interests include color imaging, color halftoning, and mul-timedia security. Dr. Oztan is a member of SPIE, IS&T, the IEEE,and the IEEE Signal Processing Society. He serves as a reviewerfor IEEE Transactions on Image Processing, IEEE Transactions onInformation Forensics and Security, IEEE Signal Processing Letters,and the SPIE/IS&T Journal of Electronic Imaging.

Gaurav Sharma is an associate professorwith the University of Rochester in the de-partment of Electrical and Computer En-gineering and in the Department of Bio-statistics and Computational Biology. He isalso the director for the Center for Emerg-ing and Innovative Sciences (CEIS), a NewYork state funded center for promoting jointuniversity-industry research and technologydevelopment, housed at the University ofRochester. He received his BE degree in

electronics and communication engineering from the Indian Instituteof Technology Roorkee (formerly the University of Roorkee) in 1990,his ME degree in electrical communication engineering from the In-dian Institute of Science, Bangalore, in 1992, and his MS degree inapplied mathematics and his PhD degree in electrical and computerengineering from North Carolina State University, Raleigh, in 1995and 1996, respectively. From August 1996 through August 2003, hewas with Xerox Research and Technology, in Webster, New York,initially as a research staff member and subsequently as a principalscientist. Dr. Sharma’s research interests include media security andwatermarking, color science and imaging, genomic signal process-ing, and image processing for visual sensor networks. He is the editorof the Color Imaging Handbook (CRC Press, 2003). He is a seniormember of the IEEE, a member of Sigma Xi, Phi Kappa Phi, Pi MuEpsilon, SPIE, IS&T, and the signal processing and communicationssocieties of the IEEE. He was the 2007 chair for the Rochester sec-tion of the IEEE and the 2003 chair for the Rochester chapter of theIEEE Signal Processing Society. He currently chairs the IEEE Sig-nal Processing Society’s Image Video and Multi-dimensional SignalProcessing (IVMSP) technical committee. He is member of the IEEESignal Processing Society’s Information Forensics and Security (IFS)technical committee and an advisory member of the IEEE StandingCommittee on Industry DSP. He is an associate editor the Journal ofElectronic Imaging and was an associate editor for the IEEE Transac-tions on Image Processing and the IEEE Transactions on InformationForensics and Security.



http://dx.doi.org/10.1117/12.179355

http://dx.doi.org/10.1117/12.179355

http://dx.doi.org/10.1364/AO.19.002513

http://dx.doi.org/10.1002/1520-6378

Documents

Per-separation clustered-dot color halftone watermarks: separation estimation based …oztanb/Papers/jei10paper.pdf · 2010-12-17 · Journal of Electronic Imaging 19(4), 043007 (Oct–Dec